Answers are grounded in this paper and cite their sources.
Deep neural networks (DNNs), renowned for their powerful representational capabilities, have achieved state-of-the-art performance in plant disease identification. However, these models are predominantly deployed in offline scenarios. In real-world applications, plant disease images with limited samples often arrive incrementally. The learning paradigm enabling models to recognize continuously emerging novel classes from limited samples without forgetting prior knowledge is termed Few-Shot Class-Incremental Learning (FSCIL). FSCIL faces significant challenges, notably catastrophic forgetting and overfitting. Given that plant disease images inherently exhibit substantial intra-class variation and minimal inter-class distinction, these challenges are particularly pronounced in plant disease recognition. To address this, we propose a Virtual class-aware Orthogonal and Contrastive Prototype calibration (VOCP) method for FSCIL-based plant disease identification. VOCP calibrates prototypes through three principles: (1) Unified class representations, (2) Prototype separation, and (3) Placeholder allocation for new classes. Specifically, we align prototypes with class centers using multi-view supervised contrastive learning, ensuring prototypes maximally represent unified class features. Concurrently, Orthogonal learning enhances prototype separation. Additionally, VOCP reserves capacity for new classes via virtual prototypes generated from predefined virtual classes. This approach effectively mitigates catastrophic forgetting and overfitting, significantly enhancing FSCIL performance for plant disease identification. Extensive experiments demonstrate VOCP's state-of-the-art results on the PlantVillage, AI Challenger 2018, and Self-Collected Plant Disease (SPD) datasets. Our method demonstrates substantial potential for advancing plant disease identification from offline to online learning paradigms.
No takes yet. Share an insight, caveat, or question.
Lin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: